Dharma AI has released the Field Extractor, a 2-billion-parameter model that pulls specified fields from document images and returns them as structured JSON, rather than the full page text that OCR tools produce. Users can either ask for everything the model can find or supply a schema naming the exact keys and value types they need. Fields that do not appear on the document come back as null instead of being guessed. The model is a LoRA adapter trained on 43,384 examples on top of DharmaOCR-LITE v1.1, the company's OCR-specialised checkpoint released in April.
Training data spans three streams: diverse English documents collected from the public web, real Brazilian Portuguese documents sourced internally, and a smaller multilingual set covering other European and Asian languages. All of it is annotated at field level and includes receipts, commercial invoices, forms, bank statements, ID documents and synthetic documents. In the company's own benchmark of 546 documents across 1,638 test cases, the model scored higher than four general-purpose frontier LLM configurations. The quality margin ranged from 1.8 points against Claude Opus 5 to 9.3 points against GPT-5.6 Luna. The company also says it costs substantially less to run than frontier APIs, although the cost comparison in the text reviewed was cut off before the full figures. The benchmark was run by the developers and has not been independently verified.
For Australian businesses, the practical change is mainly cost and workflow rather than new capability. A small model that returns clean JSON could lower the cost of processing high volumes of invoices, forms or requisitions. Many of the target documents, such as IDs and medical requisitions, contain personal information, so the Privacy Act 1988 and the Australian Privacy Principles govern how that data is sent, stored and disclosed, and health information attracts stricter handling. Teams should test on non-sensitive samples first, confirm where processing happens, and check whether a self-hosted option exists before sending customer records to any hosted service. Businesses that do not process large volumes of structured forms will see no immediate change.
Key Points
The Field Extractor returns typed JSON from document images, either open-ended or matched to a user-supplied schema, with null for fields not present.
It is a LoRA adapter on the 2B-parameter DharmaOCR-LITE v1.1, trained on 43,384 examples in English, Brazilian Portuguese and other languages.
On the company's 1,638-case benchmark, it beats four frontier LLM configurations by 1.8 to 9.3 quality points.
The company says it costs substantially less to run than frontier APIs; the full cost figures were cut off in the source text.
The results are self-reported by the developers and have not been independently verified.